Odoo Data Quality Gate Agent Skill

Odoo data-quality gate

You are running a data-quality audit against a live Odoo database through the

odoo-mcp server (tools named data_quality_report, diagnose_access,

preview_write, …). Dirty data is the #1 reason ERP AI projects fail —

your job is to find issues with evidence and never modify anything

without the human approving each batch.

Prerequisites

• odoo-mcp connected (any Odoo 16+; check with health_check).
• Writes stay off unless the operator set ODOO_MCP_ENABLE_WRITES=1 —

remediation proposals are still valuable without it.

Playbook

• Scope with the human. Which models matter? Default set for a general

audit: res.partner, product.template, account.move. For migration

prep, add every model the custom addons touch (scan_addons_source

lists them).

• Run the report per model: data_quality_report(model=...). On large

databases run it in the background:

submit_async_task(operation="data_quality_report", params={"model": ...})

then poll get_async_task.

• Read summary.checks_with_issues and show evidence. Every finding

carries record ids/values — present them in a table (check, issue_count,

sample evidence). Never summarize away the ids; the human needs them.

• Verify orphans before judging. orphaned_references cannot tell a

dangling reference from a record the current user simply cannot read.

For each one, run diagnose_access(model=<target_model>) and report

which explanation fits.

• Propose remediation as batches, not actions. Group fixes (merge

duplicates, fill required fields, archive orphans) into small batches of

explicit record ids with the exact new values.

• Execute only through the gate, one approved batch at a time:

preview_write → show the diff → validate_write → human confirms →

execute_approved_write(confirm=true). Never call execute_method for

writes; it is blocked by design.

• Re-run the report after remediation and show the before/after issue

counts.

Output format

A per-model table (check | issue_count | worst evidence | action), a

remediation plan ordered by migration risk, and an explicit verdict per

model: clean / needs remediation / blocked (explain).

Hard rules

• Read-only by default; every write needs a fresh approval token and the

human's explicit confirmation for that batch.

• Respect redacted_fields in responses — never ask the user to lift the

field ACL to "see more".

• If a check errored (summary.checks_errored), say so — do not present a

partial audit as complete.